GPT OSS Safeguard 20B
Neural Network
GPT OSS Safeguard 20B is an open 21B parameter MoE model by OpenAI for content classification and LLM output filtering.
Max answer length
(in tokens)
Context size
(in tokens)
Prompt cost
(per 1M tokens)
Answer cost
(per 1M tokens)
How it works GPT OSS Safeguard 20B?
GPT OSS Safeguard 20B is a reasoning model from OpenAI, based on gpt-oss-20b and tailored for security tasks. It features open weights and a 21B parameter Mixture-of-Experts architecture, providing low latency for content classification, LLM output filtering, and trust and safety checks. Unlike the base version, it is specialized: not a general assistant, but a policy evaluation tool. Through BotHub, you can access it from Russia without a VPN or foreign card: pay with a Russian card in rubles, only for consumed tokens that do not expire. Access over 250 models in one window, and use a unified OpenAI-compatible API to integrate checks into your product and switch models without rewriting integrations. Traffic is encrypted via AES-GCM; for companies, we offer contracts, invoices, electronic document management, and an admin panel with limits. Useful if you moderate user content, check chatbot responses before sending them to clients, sort support tickets by risk level, or test your evaluation rules on open weights.Frequently asked questions about GPT OSS Safeguard 20B
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
This is a text model with a 131,072 token context: it parses long documents, explains code, reasons through tasks, and calls external functions. The safeguard line was created for content classification based on your own rules, making it convenient for filters and stream moderation.
In a single response, the model can output up to 65,536 tokens — that's dozens of pages of text. It's enough for a large report, parsing a voluminous instruction, or detailed code analysis without splitting it into parts. If the response cuts off, just ask it to continue.
Yes, reasoning before answering is confirmed by our data: the model first builds a chain of steps and then formulates a conclusion. This is especially useful for classification tasks — you see not only the final verdict but also the logic used to reach it.
Yes. Function calling and structured JSON output are supported, so the model can be connected to your services to receive predictable fields instead of free text. In BotHub, this works via a unified OpenAI-compatible API — you can switch models without rewriting your integration.
You can input text, images, and documents — the model reads them and responds with text. There is no mention of audio or video input in our data, and image or sound generation is not supported: you receive the result in text format.
There are no specific notes about languages in our data, so we cannot promise accuracy in Russian. The most honest way is to run your real request: in BotHub, access is available without a VPN or foreign card, with payment by Russian cards based on actual token usage.